Akshay Krishnan

PhD student at GeorgiaTech. Building computational models to understand the 3D world from images.
Previously at Waymo perception.

[Sahil Khose @SahilKhose]

We are presenting WFM-Eval at two @CVPR 2026 workshops in Denver 📍

🗓️ Jun 3, Video World Models
Poster 9:50–10:40 AM, Exhibit Hall A

🗓️ Jun 4, Foundation Models Meet Embodied Agents
Poster 3:55–4:30 PM

Come say hi 👋

Work done with @AmberZhang99 @prithvijitch @judyfhoffman

[Akshay Krishnan @shay_krishnan]

@akashshrm02 @CMU_Robotics @michaelkaess @shubhtuls @GuanyaShi @mukadammh @JitendraMalikCV Congratulations Dr @akashshrm02 !🎉

[Marwan Taher @marwan_ptr]

How can we run reconstruction models like π³ and Depth Anything 3 in real-time?

We present KV-Tracker, a training-free approach, for real-time tracking of scenes and objects. Achieving up to 30 FPS!

With @alzugarayign, @makezur, @XinKong_IC and @AjdDavison

[Akshay Krishnan @shay_krishnan]

3. Bolt3D (ICCV 2025): Multi-view diffusion model that jointly generates color and 3D scene coordinates
https://t.co/7XUpM43J96

4. World-consistent Video Diffusion (CVPR 2025): Video DiT that jointly generates both color and 3D scene coordinates
https://t.co/yIE28hWKqR

[Akshay Krishnan @shay_krishnan]

It was fun presenting our work on "Orchid" - a Joint Color and Geometry Diffusion model at ICCV this year. 🌸

Orchid is a single model we trained to generate color, depth, and normals from a text prompt, all at once.

🧵 with a summary and some connections to concurrent work

[Akshay Krishnan @shay_krishnan]

1. JointDiT, (also at ICCV 2025): joint color and depth generation (no normals). Unlike Orchid, they use separate latents for color and depth.
https://t.co/BSflPJf696

2. MMGen: color, depth, normals, and also segmentations (for ImageNet classes)
https://t.co/bBP9AVhaLv

[Akshay Krishnan @shay_krishnan]

I'll be presenting OmniNOCS in the oral session 1-A at #ECCV2024 tomorrow. If you're at ECCV, do visit our poster # 183 in the first session!

[Akshay Krishnan @shay_krishnan]

New paper 📢📢

TL;DR: We introduce OmniNOCS: a multi-domain NOCS dataset with 90+ object classes. Our model trained on OmniNOCS can predict 6DoF object pose and partial shape from their 2D detections, even generalizing to in-the-wild images!

Background & details in 🧵 below

[Akshay Krishnan @shay_krishnan]

@mzubairirshad Thank you @mzubairirshad! 😊

[Akshay Krishnan @shay_krishnan]

New paper 📢📢

TL;DR: We introduce OmniNOCS: a multi-domain NOCS dataset with 90+ object classes. Our model trained on OmniNOCS can predict 6DoF object pose and partial shape from their 2D detections, even generalizing to in-the-wild images!

Background & details in 🧵 below

[Akshay Krishnan @shay_krishnan]

@_akhaliq Thanks for sharing our work @_akhaliq! I also shared a thread with more details :) https://t.co/KGniTnrhgR

[Akshay Krishnan @shay_krishnan]

New paper 📢📢

TL;DR: We introduce OmniNOCS: a multi-domain NOCS dataset with 90+ object classes. Our model trained on OmniNOCS can predict 6DoF object pose and partial shape from their 2D detections, even generalizing to in-the-wild images!

Background & details in 🧵 below

[Akshay Krishnan @shay_krishnan]

OmniNOCS has been accepted to ECCV 24! 🥳
Please consider using it to train or evaluate models for 3D object coordinate regression.

Joint work with my awesome collaborators @_abhijit_kundu_, @kmaninis, Matthew Brown at @GoogleDeepMind and @jhhays at @GeorgiaTech @mlatgt .

[Akshay Krishnan @shay_krishnan]

We trained a new model (NOCSformer) on OmniNOCS to predict the NOCS and pose of objects from RGB alone (no depth input!). The large-scale training allows it to generalize to in-the-wild images, including objects from COCO.

Data, paper, more results at: https://t.co/5mDKqH3CFO

[Akshay Krishnan @shay_krishnan]

@nice_byte The latter, I do think the process significantly helped me though.

[Akshay Krishnan @shay_krishnan]

@nice_byte I had the readability process at waymo :)